Re: [PATCH] iio: bu27008: Add processed illuminance channel

[Matti Vaittinen <mazziesaccount@gmail.com>]

On 10/10/23 12:40, Jonathan Cameron wrote:
> On Fri, 6 Oct 2023 08:01:15 +0300
> Matti Vaittinen <mazziesaccount@gmail.com> wrote:
> 
>> On 10/5/23 18:14, Jonathan Cameron wrote:
>>> On Mon, 2 Oct 2023 15:33:41 +0300
>>> Matti Vaittinen <mazziesaccount@gmail.com> wrote:
>>>    
>>>> The RGB + IR data can be used to calculate illuminance value (Luxes).
>>>> Implement the equation obtained from the ROHM HW colleagues and add a
>>>> light data channel outputting illuminance values in (nano) Luxes.
>>> Units in the ABI doc for illuminance are Lux, not nanolux.
>>> I'm guessing that you actually provide it in Lux but via scale.
>>>
>>> Make that clearer in this description if so.
>>
>> Yep. Also, the "processed" is misleading as I implement a raw channel. I
>> did originally think I'll only implement the read_raw (as I thought
>> we'll need all RGBC + IR and end up doing two accesses - which wouldn't
>> be nice due to the doubled measurement time). I actually did that and
>> used INT_PLUS_NANO. While implementing this I noticed the 'clear' data
>> was not used - and thought I might as well support buffering when RGB+IR
>> are enabled. I needed the scale to get the buffered values to decent
>> format though - so I converted channel to raw one and added scale. The
>> commit title still contains the 'processed' which reflects the original
>> thinking. Thanks for pointing out the confusion.
>>
>>>> Both the read_raw and buffering values is supported, with the limitation
>>>> that buffering is only allowed when suitable scan-mask is used. (RGB+IR,
>>>> no clear).
>>>>
>>>> The equation has been developed by ROHM HW colleagues for open air sensor.
>>>> Adding any lens to the sensor is likely to impact to the used c1, c2, c3
>>>> coefficients. Also, The output values have only been tested on BU27008.
>>>>
>>>> According to the HW colleagues, the very same equation should work also
>>>> on BU27010.
>>>>
>>>> Calculate and output illuminance values from BU27008 and BU27010.
>>>>
>>>> Signed-off-by: Matti Vaittinen <mazziesaccount@gmail.com>
>>>>   
>>>
>>> A few comments inline, but in general looks fine to me.
>>
>> Thanks Jonathan. I had to give also the BU27008 sensor away for a while.
>> I guess I won't send the next version until I am able to do some very
>> basic testing even if the changes were minor. That's probably sometime
>> next week.
>>
>>>
>>> Jonathan
>>>    
>>>> ---
>>>>

//snip

>>>>    
>>>> +static int bu27008_get_rgb_ir(struct bu27008_data *data, unsigned int *red,
>>>> +		    unsigned int *green, unsigned int *blue, unsigned int *ir)
>>>> +{
>>>> +	int ret, chan_sel, int_time, tmpret, valid;
>>>> +	__le16 chans[BU27008_NUM_HW_CHANS];
>>>> +
>>>> +	chan_sel = BU27008_BLUE2_IR3 << (ffs(data->cd->chan_sel_mask) - 1);
>>>> +
>>>> +	ret = regmap_update_bits(data->regmap, data->cd->chan_sel_reg,
>>>> +				 data->cd->chan_sel_mask, chan_sel);
>>>> +	if (ret)
>>>> +		return ret;
>>>> +
>>>> +	ret = bu27008_meas_set(data, true);
>>>> +	if (ret)
>>>> +		return ret;
>>>> +
>>>> +	ret = bu27008_get_int_time_us(data);
>>>> +	if (ret < 0)
>>>> +		int_time = BU27008_MEAS_TIME_MAX_MS;
>>>> +	else
>>>> +		int_time = ret / USEC_PER_MSEC;
>>>> +
>>>> +	msleep(int_time);
>>>> +
>>>> +	ret = regmap_read_poll_timeout(data->regmap, data->cd->valid_reg,
>>>> +				       valid, (valid & BU27008_MASK_VALID),
>>>> +				       BU27008_VALID_RESULT_WAIT_QUANTA_US,
>>>> +				       BU27008_MAX_VALID_RESULT_WAIT_US);
>>>> +	if (ret)
>>>> +		goto out;
>>>> +
>>>> +	ret = regmap_bulk_read(data->regmap, BU27008_REG_DATA0_LO, chans,
>>>> +			       sizeof(chans));
>>>> +	if (ret)
>>>> +		goto out;
>>>> +
>>>> +	*red = le16_to_cpu(chans[0]);
>>>> +	*green = le16_to_cpu(chans[1]);
>>>> +	*blue = le16_to_cpu(chans[2]);
>>>> +	*ir = le16_to_cpu(chans[3]);
>>>
>>> I'd be tempted to use an array + definitely pass them as u16 rather
>>> than unsigned int.
>>
>> I'm not really convinced the u16 is better here. We need the 32 bits
>> later for the calculations - and (afaics) using natural size int for
>> arguments shouldn't harm. We read the channel data to correct type array
>> so code should be pretty clear as to what we have in HW.
> 
> ok.  I don't like lack of range clamping - so at the point of the caller
> I can't immediately see that these will be sub 16bit value.  Not htat
> important I guess.
> 
>>
>> Also, I think that having an array obfuscates what element is which
>> channel because these ICs didn't have the 1 to 1 mapping from channel
>> index to colour. I was thinking of adding a struct for this but decided
>> to just keep it simple and clear.
> A struct or array + enum would work.
> I just don't much like lots of very similar parameters.

Right. A struct is not a problem. I am less fond with an enum because 
the HW channel which carries a specific color may change. I think adding 
an enum which indicates a place where a color is in array may be 
misleading one to think that HW has a fixed channel (data register 
address) for a colour. (I think that is quite usual while ICs where one 
color may be in different channels depending on the config are likely to 
be rare... I haven't read too many light sensor specs/drivers to know 
for sure though).

>>
>>>> +
>>>> +out:
>>>> +	tmpret = bu27008_meas_set(data, false);
>>>> +	if (tmpret)
>>>> +		dev_warn(data->dev, "Stopping measurement failed\n");
>>>> +
>>>> +	return ret;
>>>> +}
>>>> +
>>>> +/*
>>>> + * Following equation for computing lux out of register values was given by
>>>> + * ROHM HW colleagues;
>>>> + *
>>>> + * Red = RedData*1024 / Gain * 20 / meas_mode
>>>> + * Green = GreenData* 1024 / Gain * 20 / meas_mode
>>>> + * Blue = BlueData* 1024 / Gain * 20 / meas_mode
>>>> + * IR = IrData* 1024 / Gain * 20 / meas_mode
>>>> + *
>>>> + * where meas_mode is the integration time in mS / 10
>>>> + *
>>>> + * IRratio = (IR > 0.18 * Green) ? 0 : 1
>>>> + *
>>>> + * Lx = max(c1*Red + c2*Green + c3*Blue,0)
>>>> + *
>>>> + * for
>>>> + * IRratio 0: c1 = -0.00002237, c2 = 0.0003219, c3 = -0.000120371
>>>> + * IRratio 1: c1 = -0.00001074, c2 = 0.000305415, c3 = -0.000129367
>>>> + */
>>>> +
>>>> +/*
>>>> + * The max chan data is 0xffff. When we multiply it by 1024 * 20, we'll get
>>>> + * 0x4FFFB000 which still fits in 32-bit integer. So this can't overflow.
>>>> + */
>>>> +#define NORM_CHAN_DATA_FOR_LX_CALC(chan, gain, time) ((chan) * 1024 * 20 / \
>>>> +				   (gain) / (time))
>>>> +static u64 bu27008_calc_nlux(struct bu27008_data *data, unsigned int red,
>>>> +		unsigned int green, unsigned int blue,  unsigned int ir,
>>>> +		unsigned int gain, unsigned int gain_ir, unsigned int time)
>>>> +{
>>>> +	s64 c1, c2, c3, nlux;
>>>> +
>>>> +	time /= 10000;
>>>> +	ir = NORM_CHAN_DATA_FOR_LX_CALC(ir, gain_ir, time);
>>>> +	red = NORM_CHAN_DATA_FOR_LX_CALC(red, gain, time);
>>>> +	green = NORM_CHAN_DATA_FOR_LX_CALC(green, gain, time);
>>>> +	blue = NORM_CHAN_DATA_FOR_LX_CALC(blue, gain, time);
>>
>>> I'd prefer to see the inputs parameters and the normalized version given different
>>> names. Also the inputs are still u16, so nice to reflect that here.
>>
>> So, you suggest we bring the data as u16 until here and only here we
>> assign it into 32bit variables when doing the 'normalization'? I'm sure
>> it works, but I dislike doing computations like multiplying u16 by u32
>> as I never know (out of my head) how the implicit type conversions work
>> and if we get some results cropped. Adding the casts to computation make
>> it less pretty for my eyes while having all variables in large enough
>> types does not leave me wondering if it works correctly and if explicit
>> casts are needed.
>>
>> I am not strongly opposing this though if you insist - I am sure I can
>> at the end of the day get the code right - but I am afraid I will later
>> look at the code and wonder if it contains hideous issues...
> 
> This isn't particularly important either way.  My gut would have
> been to keep them as __le16 to the point where the maths happens
> but I don't mind it happening elsewhere.
> 
> I do want different names though given the inputs and outputs are
> different 'things'.
> 

I'll try improving this and hopefully send a new version later this 
week. I should get the sensor at my hands latest at Thursday.

> 
>>
>>> Also when doing normalization I'd used fixed with types so there is no
>>> confusion over what was intended (here u32)
>>
>> Ok.
>>
>>>    
>>>> +
>>>> +	if ((u64)ir * 100LLU > 18LLU * (u64)green) {
>>>
>>> Putting scaling for ir to the right and green to the left is
>>> unusual. I'd chose one side and stick to it.
>>
>> Sorry Jonathan. I must be a bit slow today but I just seem to not be
>> able to think how you would like to have this? I think this line is
>> somehow mappable to the:
> 
> if ((u64)ir * 100LLU > (u64)green * 18LLU)
> or
> if ((100LLU * (u64)ir > 18LLU * (u64)green)
> 
> Either is fine.  Just don't like the scaling from different sides of
> the variable.  I can see how you got there from 0.18 * Green but equally
> valid to premultiply by 100 as it is to post multiply (when doing the
> maths on paper).

Now I see what you meant :) I misread your comment to meant that you 
didn't like the scaling on both sides of the '>'. /me feels slightly stupid.

Thanks again!

>>
>> IRratio = (IR > 0.18 * Green) ? 0 : 1
>> formula I got from HW colleagues and added in the comment preceding the
>> function.
>>

Yours,
	-- Matti

-- 
Matti Vaittinen
Linux kernel developer at ROHM Semiconductors
Oulu Finland

~~ When things go utterly wrong vim users can always type :help! ~~